Shader interpreter causing a crash on AMD GPU

[jackie1218]

Quick summary

When the options Async + Shader interpreter, or Shader interpreter only are used, RPCS3 crashes and closes. If the shader cache is partially built for a game, the first time the interpreter is called, RPCS3 will crash.

Details

No response

Attach a log file

RPCS3.log

Attach capture files for visual issues

No response

System configuration

Windows 11 24H2
AMD Ryzen 7 5700X3D
AMD Radeon RX 7900 XT (driver version 24.10.1)

Other details

No response

[jackie1218]

This issue is still present on 24.12.1 (which came out literally minutes after me writing this issue)

[Valtekken]

This issue is still present on 24.12.1 (which came out literally minutes after me writing this issue)

It was actually released yesterday https://videocardz.com/driver/amd-radeon-software-adrenalin-24-12-1

[jackie1218]

This crash wasn't fixed by #16540, still happens in latest master

[kd-11]

Need to validate the shader and send the spv off to amd to fix the compiler if it's broken.

[Darkhost1999]

Are there specific games to trigger the crash?
RPCS3.log
Vulkan:
Adapter: AMD Radeon(TM) Graphics

When the option Async + Shader interpreter is used everything is fine in my game.
When the option Shader interpreter only is used more than half the cutscenes and all the gameplay is a black screen with only subtitles and HUD showing regardless of AMD vs Nvidia

[kd-11]

Visual glitches are fine and expected. 99% of them are because we do not emulate the RSX weird floating point behavior in interpreter mode for performance reasons.
The shader compiler crashes are annoying though. I've seen this a few times when debugging.
The interpreter shader also takes like a full minute to build while its only a few seconds on NV drivers.
NV vulkan driver used to be pretty bad but they overhauled their shader compiler a few years ago and bow it's pretty good. No crashes so far (from shader compiler, device lost mess is still a problem with NV)

[jackie1218]

Are there specific games to trigger the crash?

No, I can reliably trigger it in any game I try. RPCS3 just closes as soon as the shader interpreter is used. If you already have a shader cache for a game, you'd have to clear it before you'd see this crash (or get to a part of the game that forces you to compile them).

[kd-11]

@Jacoby1218 Do you mean it actually crashes each and every attempt? I've had the interpreter in use over 3 generations of AMD cards (polaris, RDNA2 and RDNA3) and it usually works around 80% of the time on my setup. I can sometimes see crashes but it is rare and the type is always access violation in event viewer logs. Typically if the first interpreter shader compiles the rest compile fine since the code is mostly the same and only state is changing.
If you have 100% crash rate that could be something more useful to report.

[kd-11]

Ok, I just tried again, shader interpreter only, RDNA3 all caches deleted and it started happening reliably at boot. I'll run this through spirv-val, let's see if there's any bugs in the shader or if I just need to file this with AMD.

[kd-11]

Hmm, looks like it's our bug. I need to reimplement the binding slots stuff a bit.

[kd-11]

Should be fixed by #16576
Test using DeS on RDNA3 (you can see broken shadow caused by interpreter):

I also fixed the black textures in most games when using interpreter. That was a regression.